Weights & Biases

Weights & Biases documentation Example Usage

What it is¶

Weights & Biases (W&B) is a cloud-hosted experiment tracker. Results are stored at wandb.ai and every run gets a shareable URL so there is no local infrastructure required for anyone who needs to view them.

Why use it¶

W&B takes the same idea as MLflow: logs parameters, metrics, and artifacts per run and makes the results accessible from anywhere via a browser link. That makes it the natural choice for distributed teams: a collaborator clicks a URL and sees the dashboards, no local setup, no shared filesystem.

When to use it¶

• Team collaboration with remote members who should not need local setup.

• When run URLs need to live in issues, slides, or papers.

• When you want the built-in dashboards without wiring up your own.

• When storing run data on W&B servers is acceptable.

How to use it¶

Install¶

uv sync   # wandb is in pyproject.toml
# or: uv add wandb

Authenticate (one-time)¶

wandb login
# Stores the API key in ~/.netrc

For CI or containers, set the key as an environment variable instead:

export WANDB_API_KEY=your_api_key_here

Configure with environment variables¶

Run and track¶

python scripts/train_sst_wandb.py --run-name "experiment_1"
# → Run URL printed to console

What gets tracked per run¶

Each run in workshop-sst logs:

Load a registered model¶

import joblib
from pathlib import Path
import wandb

run = wandb.init()
artifact = run.use_artifact("wandb-registry-workshop-sst/models:v0.6.2")
model = joblib.load(Path(artifact.download()) / "model.joblib")

Run offline on HPC¶

If the cluster has no outbound network access:

WANDB_MODE=offline python scripts/train_sst_wandb.py

Sync the runs when network is available:

wandb sync wandb/offline-run-*

Notes from use¶

A few things worth knowing before you wire W&B into a training script:

• Package version does not log automatically. If you want the library version alongside the run, pipe it through wandb.config explicitly (e.g. wandb.config["pkg_version"] = importlib.metadata.version("your-pkg")).

• Metric charts stay empty for single-point runs. W&B’s line charts expect a series. If your script only logs a final R² and RMSE once per run, the per-metric chart will appear empty. Log per-epoch or per-step values to populate the chart, or rely on the project-level bar chart instead.

• The project-level bar chart of metrics is easy to miss in the UI. Look for it on the project’s main dashboard rather than inside an individual run.

Pros and cons¶

Reference¶

• Weights & Biases documentation

• W&B Artifacts

• W&B Model Registry

• Offline mode

• W&B pairs with DVC the same way MLflow does — see MLflow + DVC and substitute W&B for the tracker.